fs/nfs/pagelist.c

   1 // SPDX-License-Identifier: GPL-2.0-only
   2 /*
   3  * linux/fs/nfs/pagelist.c
   4  *
   5  * A set of helper functions for managing NFS read and write requests.
   6  * The main purpose of these routines is to provide support for the
   7  * coalescing of several requests into a single RPC call.
   8  *
   9  * Copyright 2000, 2001 (c) Trond Myklebust <trond.myklebust@fys.uio.no>
  10  *
  11  */
  12
  13 #include <linux/slab.h>
  14 #include <linux/file.h>
  15 #include <linux/sched.h>
  16 #include <linux/sunrpc/clnt.h>
  17 #include <linux/nfs.h>
  18 #include <linux/nfs3.h>
  19 #include <linux/nfs4.h>
  20 #include <linux/nfs_fs.h>
  21 #include <linux/nfs_page.h>
  22 #include <linux/nfs_mount.h>
  23 #include <linux/export.h>
  24
  25 #include "internal.h"
  26 #include "pnfs.h"
  27
  28 #define NFSDBG_FACILITY         NFSDBG_PAGECACHE
  29
  30 static struct kmem_cache *nfs_page_cachep;
  31 static const struct rpc_call_ops nfs_pgio_common_ops;
  32
  33 struct nfs_pgio_mirror *
  34 nfs_pgio_current_mirror(struct nfs_pageio_descriptor *desc)
  35 {
  36         return &desc->pg_mirrors[desc->pg_mirror_idx];
  37 }
  38 EXPORT_SYMBOL_GPL(nfs_pgio_current_mirror);
  39
  40 void nfs_pgheader_init(struct nfs_pageio_descriptor *desc,
  41                        struct nfs_pgio_header *hdr,
  42                        void (*release)(struct nfs_pgio_header *hdr))
  43 {
  44         struct nfs_pgio_mirror *mirror = nfs_pgio_current_mirror(desc);
  45
  46
  47         hdr->req = nfs_list_entry(mirror->pg_list.next);
  48         hdr->inode = desc->pg_inode;
  49         hdr->cred = nfs_req_openctx(hdr->req)->cred;
  50         hdr->io_start = req_offset(hdr->req);
  51         hdr->good_bytes = mirror->pg_count;
  52         hdr->io_completion = desc->pg_io_completion;
  53         hdr->dreq = desc->pg_dreq;
  54         hdr->release = release;
  55         hdr->completion_ops = desc->pg_completion_ops;
  56         if (hdr->completion_ops->init_hdr)
  57                 hdr->completion_ops->init_hdr(hdr);
  58
  59         hdr->pgio_mirror_idx = desc->pg_mirror_idx;
  60 }
  61 EXPORT_SYMBOL_GPL(nfs_pgheader_init);
  62
  63 void nfs_set_pgio_error(struct nfs_pgio_header *hdr, int error, loff_t pos)
  64 {
  65         unsigned int new = pos - hdr->io_start;
  66
  67         if (hdr->good_bytes > new) {
  68                 hdr->good_bytes = new;
  69                 clear_bit(NFS_IOHDR_EOF, &hdr->flags);
  70                 if (!test_and_set_bit(NFS_IOHDR_ERROR, &hdr->flags))
  71                         hdr->error = error;
  72         }
  73 }
  74
  75 static inline struct nfs_page *
  76 nfs_page_alloc(void)
  77 {
  78         struct nfs_page *p = kmem_cache_zalloc(nfs_page_cachep, GFP_KERNEL);
  79         if (p)
  80                 INIT_LIST_HEAD(&p->wb_list);
  81         return p;
  82 }
  83
  84 static inline void
  85 nfs_page_free(struct nfs_page *p)
  86 {
  87         kmem_cache_free(nfs_page_cachep, p);
  88 }
  89
  90 /**
  91  * nfs_iocounter_wait - wait for i/o to complete
  92  * @l_ctx: nfs_lock_context with io_counter to use
  93  *
  94  * returns -ERESTARTSYS if interrupted by a fatal signal.
  95  * Otherwise returns 0 once the io_count hits 0.
  96  */
  97 int
  98 nfs_iocounter_wait(struct nfs_lock_context *l_ctx)
  99 {
 100         return wait_var_event_killable(&l_ctx->io_count,
 101                                        !atomic_read(&l_ctx->io_count));
 102 }
 103
 104 /**
 105  * nfs_async_iocounter_wait - wait on a rpc_waitqueue for I/O
 106  * to complete
 107  * @task: the rpc_task that should wait
 108  * @l_ctx: nfs_lock_context with io_counter to check
 109  *
 110  * Returns true if there is outstanding I/O to wait on and the
 111  * task has been put to sleep.
 112  */
 113 bool
 114 nfs_async_iocounter_wait(struct rpc_task *task, struct nfs_lock_context *l_ctx)
 115 {
 116         struct inode *inode = d_inode(l_ctx->open_context->dentry);
 117         bool ret = false;
 118
 119         if (atomic_read(&l_ctx->io_count) > 0) {
 120                 rpc_sleep_on(&NFS_SERVER(inode)->uoc_rpcwaitq, task, NULL);
 121                 ret = true;
 122         }
 123
 124         if (atomic_read(&l_ctx->io_count) == 0) {
 125                 rpc_wake_up_queued_task(&NFS_SERVER(inode)->uoc_rpcwaitq, task);
 126                 ret = false;
 127         }
 128
 129         return ret;
 130 }
 131 EXPORT_SYMBOL_GPL(nfs_async_iocounter_wait);
 132
 133 /*
 134  * nfs_page_lock_head_request - page lock the head of the page group
 135  * @req: any member of the page group
 136  */
 137 struct nfs_page *
 138 nfs_page_group_lock_head(struct nfs_page *req)
 139 {
 140         struct nfs_page *head = req->wb_head;
 141
 142         while (!nfs_lock_request(head)) {
 143                 int ret = nfs_wait_on_request(head);
 144                 if (ret < 0)
 145                         return ERR_PTR(ret);
 146         }
 147         if (head != req)
 148                 kref_get(&head->wb_kref);
 149         return head;
 150 }
 151
 152 /*
 153  * nfs_unroll_locks -  unlock all newly locked reqs and wait on @req
 154  * @head: head request of page group, must be holding head lock
 155  * @req: request that couldn't lock and needs to wait on the req bit lock
 156  *
 157  * This is a helper function for nfs_lock_and_join_requests
 158  * returns 0 on success, < 0 on error.
 159  */
 160 static void
 161 nfs_unroll_locks(struct nfs_page *head, struct nfs_page *req)
 162 {
 163         struct nfs_page *tmp;
 164
 165         /* relinquish all the locks successfully grabbed this run */
 166         for (tmp = head->wb_this_page ; tmp != req; tmp = tmp->wb_this_page) {
 167                 if (!kref_read(&tmp->wb_kref))
 168                         continue;
 169                 nfs_unlock_and_release_request(tmp);
 170         }
 171 }
 172
 173 /*
 174  * nfs_page_group_lock_subreq -  try to lock a subrequest
 175  * @head: head request of page group
 176  * @subreq: request to lock
 177  *
 178  * This is a helper function for nfs_lock_and_join_requests which
 179  * must be called with the head request and page group both locked.
 180  * On error, it returns with the page group unlocked.
 181  */
 182 static int
 183 nfs_page_group_lock_subreq(struct nfs_page *head, struct nfs_page *subreq)
 184 {
 185         int ret;
 186
 187         if (!kref_get_unless_zero(&subreq->wb_kref))
 188                 return 0;
 189         while (!nfs_lock_request(subreq)) {
 190                 nfs_page_group_unlock(head);
 191                 ret = nfs_wait_on_request(subreq);
 192                 if (!ret)
 193                         ret = nfs_page_group_lock(head);
 194                 if (ret < 0) {
 195                         nfs_unroll_locks(head, subreq);
 196                         nfs_release_request(subreq);
 197                         return ret;
 198                 }
 199         }
 200         return 0;
 201 }
 202
 203 /*
 204  * nfs_page_group_lock_subrequests -  try to lock the subrequests
 205  * @head: head request of page group
 206  *
 207  * This is a helper function for nfs_lock_and_join_requests which
 208  * must be called with the head request locked.
 209  */
 210 int nfs_page_group_lock_subrequests(struct nfs_page *head)
 211 {
 212         struct nfs_page *subreq;
 213         int ret;
 214
 215         ret = nfs_page_group_lock(head);
 216         if (ret < 0)
 217                 return ret;
 218         /* lock each request in the page group */
 219         for (subreq = head->wb_this_page; subreq != head;
 220                         subreq = subreq->wb_this_page) {
 221                 ret = nfs_page_group_lock_subreq(head, subreq);
 222                 if (ret < 0)
 223                         return ret;
 224         }
 225         nfs_page_group_unlock(head);
 226         return 0;
 227 }
 228
 229 /*
 230  * nfs_page_set_headlock - set the request PG_HEADLOCK
 231  * @req: request that is to be locked
 232  *
 233  * this lock must be held when modifying req->wb_head
 234  *
 235  * return 0 on success, < 0 on error
 236  */
 237 int
 238 nfs_page_set_headlock(struct nfs_page *req)
 239 {
 240         if (!test_and_set_bit(PG_HEADLOCK, &req->wb_flags))
 241                 return 0;
 242
 243         set_bit(PG_CONTENDED1, &req->wb_flags);
 244         smp_mb__after_atomic();
 245         return wait_on_bit_lock(&req->wb_flags, PG_HEADLOCK,
 246                                 TASK_UNINTERRUPTIBLE);
 247 }
 248
 249 /*
 250  * nfs_page_clear_headlock - clear the request PG_HEADLOCK
 251  * @req: request that is to be locked
 252  */
 253 void
 254 nfs_page_clear_headlock(struct nfs_page *req)
 255 {
 256         smp_mb__before_atomic();
 257         clear_bit(PG_HEADLOCK, &req->wb_flags);
 258         smp_mb__after_atomic();
 259         if (!test_bit(PG_CONTENDED1, &req->wb_flags))
 260                 return;
 261         wake_up_bit(&req->wb_flags, PG_HEADLOCK);
 262 }
 263
 264 /*
 265  * nfs_page_group_lock - lock the head of the page group
 266  * @req: request in group that is to be locked
 267  *
 268  * this lock must be held when traversing or modifying the page
 269  * group list
 270  *
 271  * return 0 on success, < 0 on error
 272  */
 273 int
 274 nfs_page_group_lock(struct nfs_page *req)
 275 {
 276         int ret;
 277
 278         ret = nfs_page_set_headlock(req);
 279         if (ret || req->wb_head == req)
 280                 return ret;
 281         return nfs_page_set_headlock(req->wb_head);
 282 }
 283
 284 /*
 285  * nfs_page_group_unlock - unlock the head of the page group
 286  * @req: request in group that is to be unlocked
 287  */
 288 void
 289 nfs_page_group_unlock(struct nfs_page *req)
 290 {
 291         if (req != req->wb_head)
 292                 nfs_page_clear_headlock(req->wb_head);
 293         nfs_page_clear_headlock(req);
 294 }
 295
 296 /*
 297  * nfs_page_group_sync_on_bit_locked
 298  *
 299  * must be called with page group lock held
 300  */
 301 static bool
 302 nfs_page_group_sync_on_bit_locked(struct nfs_page *req, unsigned int bit)
 303 {
 304         struct nfs_page *head = req->wb_head;
 305         struct nfs_page *tmp;
 306
 307         WARN_ON_ONCE(!test_bit(PG_HEADLOCK, &head->wb_flags));
 308         WARN_ON_ONCE(test_and_set_bit(bit, &req->wb_flags));
 309
 310         tmp = req->wb_this_page;
 311         while (tmp != req) {
 312                 if (!test_bit(bit, &tmp->wb_flags))
 313                         return false;
 314                 tmp = tmp->wb_this_page;
 315         }
 316
 317         /* true! reset all bits */
 318         tmp = req;
 319         do {
 320                 clear_bit(bit, &tmp->wb_flags);
 321                 tmp = tmp->wb_this_page;
 322         } while (tmp != req);
 323
 324         return true;
 325 }
 326
 327 /*
 328  * nfs_page_group_sync_on_bit - set bit on current request, but only
 329  *   return true if the bit is set for all requests in page group
 330  * @req - request in page group
 331  * @bit - PG_* bit that is used to sync page group
 332  */
 333 bool nfs_page_group_sync_on_bit(struct nfs_page *req, unsigned int bit)
 334 {
 335         bool ret;
 336
 337         nfs_page_group_lock(req);
 338         ret = nfs_page_group_sync_on_bit_locked(req, bit);
 339         nfs_page_group_unlock(req);
 340
 341         return ret;
 342 }
 343
 344 /*
 345  * nfs_page_group_init - Initialize the page group linkage for @req
 346  * @req - a new nfs request
 347  * @prev - the previous request in page group, or NULL if @req is the first
 348  *         or only request in the group (the head).
 349  */
 350 static inline void
 351 nfs_page_group_init(struct nfs_page *req, struct nfs_page *prev)
 352 {
 353         struct inode *inode;
 354         WARN_ON_ONCE(prev == req);
 355
 356         if (!prev) {
 357                 /* a head request */
 358                 req->wb_head = req;
 359                 req->wb_this_page = req;
 360         } else {
 361                 /* a subrequest */
 362                 WARN_ON_ONCE(prev->wb_this_page != prev->wb_head);
 363                 WARN_ON_ONCE(!test_bit(PG_HEADLOCK, &prev->wb_head->wb_flags));
 364                 req->wb_head = prev->wb_head;
 365                 req->wb_this_page = prev->wb_this_page;
 366                 prev->wb_this_page = req;
 367
 368                 /* All subrequests take a ref on the head request until
 369                  * nfs_page_group_destroy is called */
 370                 kref_get(&req->wb_head->wb_kref);
 371
 372                 /* grab extra ref and bump the request count if head request
 373                  * has extra ref from the write/commit path to handle handoff
 374                  * between write and commit lists. */
 375                 if (test_bit(PG_INODE_REF, &prev->wb_head->wb_flags)) {
 376                         inode = page_file_mapping(req->wb_page)->host;
 377                         set_bit(PG_INODE_REF, &req->wb_flags);
 378                         kref_get(&req->wb_kref);
 379                         atomic_long_inc(&NFS_I(inode)->nrequests);
 380                 }
 381         }
 382 }
 383
 384 /*
 385  * nfs_page_group_destroy - sync the destruction of page groups
 386  * @req - request that no longer needs the page group
 387  *
 388  * releases the page group reference from each member once all
 389  * members have called this function.
 390  */
 391 static void
 392 nfs_page_group_destroy(struct kref *kref)
 393 {
 394         struct nfs_page *req = container_of(kref, struct nfs_page, wb_kref);
 395         struct nfs_page *head = req->wb_head;
 396         struct nfs_page *tmp, *next;
 397
 398         if (!nfs_page_group_sync_on_bit(req, PG_TEARDOWN))
 399                 goto out;
 400
 401         tmp = req;
 402         do {
 403                 next = tmp->wb_this_page;
 404                 /* unlink and free */
 405                 tmp->wb_this_page = tmp;
 406                 tmp->wb_head = tmp;
 407                 nfs_free_request(tmp);
 408                 tmp = next;
 409         } while (tmp != req);
 410 out:
 411         /* subrequests must release the ref on the head request */
 412         if (head != req)
 413                 nfs_release_request(head);
 414 }
 415
 416 static struct nfs_page *
 417 __nfs_create_request(struct nfs_lock_context *l_ctx, struct page *page,
 418                    unsigned int pgbase, unsigned int offset,
 419                    unsigned int count)
 420 {
 421         struct nfs_page         *req;
 422         struct nfs_open_context *ctx = l_ctx->open_context;
 423
 424         if (test_bit(NFS_CONTEXT_BAD, &ctx->flags))
 425                 return ERR_PTR(-EBADF);
 426         /* try to allocate the request struct */
 427         req = nfs_page_alloc();
 428         if (req == NULL)
 429                 return ERR_PTR(-ENOMEM);
 430
 431         req->wb_lock_context = l_ctx;
 432         refcount_inc(&l_ctx->count);
 433         atomic_inc(&l_ctx->io_count);
 434
 435         /* Initialize the request struct. Initially, we assume a
 436          * long write-back delay. This will be adjusted in
 437          * update_nfs_request below if the region is not locked. */
 438         req->wb_page    = page;
 439         if (page) {
 440                 req->wb_index = page_index(page);
 441                 get_page(page);
 442         }
 443         req->wb_offset  = offset;
 444         req->wb_pgbase  = pgbase;
 445         req->wb_bytes   = count;
 446         kref_init(&req->wb_kref);
 447         req->wb_nio = 0;
 448         return req;
 449 }
 450
 451 /**
 452  * nfs_create_request - Create an NFS read/write request.
 453  * @ctx: open context to use
 454  * @page: page to write
 455  * @offset: starting offset within the page for the write
 456  * @count: number of bytes to read/write
 457  *
 458  * The page must be locked by the caller. This makes sure we never
 459  * create two different requests for the same page.
 460  * User should ensure it is safe to sleep in this function.
 461  */
 462 struct nfs_page *
 463 nfs_create_request(struct nfs_open_context *ctx, struct page *page,
 464                    unsigned int offset, unsigned int count)
 465 {
 466         struct nfs_lock_context *l_ctx = nfs_get_lock_context(ctx);
 467         struct nfs_page *ret;
 468
 469         if (IS_ERR(l_ctx))
 470                 return ERR_CAST(l_ctx);
 471         ret = __nfs_create_request(l_ctx, page, offset, offset, count);
 472         if (!IS_ERR(ret))
 473                 nfs_page_group_init(ret, NULL);
 474         nfs_put_lock_context(l_ctx);
 475         return ret;
 476 }
 477
 478 static struct nfs_page *
 479 nfs_create_subreq(struct nfs_page *req,
 480                   unsigned int pgbase,
 481                   unsigned int offset,
 482                   unsigned int count)
 483 {
 484         struct nfs_page *last;
 485         struct nfs_page *ret;
 486
 487         ret = __nfs_create_request(req->wb_lock_context, req->wb_page,
 488                         pgbase, offset, count);
 489         if (!IS_ERR(ret)) {
 490                 /* find the last request */
 491                 for (last = req->wb_head;
 492                      last->wb_this_page != req->wb_head;
 493                      last = last->wb_this_page)
 494                         ;
 495
 496                 nfs_lock_request(ret);
 497                 ret->wb_index = req->wb_index;
 498                 nfs_page_group_init(ret, last);
 499                 ret->wb_nio = req->wb_nio;
 500         }
 501         return ret;
 502 }
 503
 504 /**
 505  * nfs_unlock_request - Unlock request and wake up sleepers.
 506  * @req: pointer to request
 507  */
 508 void nfs_unlock_request(struct nfs_page *req)
 509 {
 510         if (!NFS_WBACK_BUSY(req)) {
 511                 printk(KERN_ERR "NFS: Invalid unlock attempted\n");
 512                 BUG();
 513         }
 514         smp_mb__before_atomic();
 515         clear_bit(PG_BUSY, &req->wb_flags);
 516         smp_mb__after_atomic();
 517         if (!test_bit(PG_CONTENDED2, &req->wb_flags))
 518                 return;
 519         wake_up_bit(&req->wb_flags, PG_BUSY);
 520 }
 521
 522 /**
 523  * nfs_unlock_and_release_request - Unlock request and release the nfs_page
 524  * @req: pointer to request
 525  */
 526 void nfs_unlock_and_release_request(struct nfs_page *req)
 527 {
 528         nfs_unlock_request(req);
 529         nfs_release_request(req);
 530 }
 531
 532 /*
 533  * nfs_clear_request - Free up all resources allocated to the request
 534  * @req:
 535  *
 536  * Release page and open context resources associated with a read/write
 537  * request after it has completed.
 538  */
 539 static void nfs_clear_request(struct nfs_page *req)
 540 {
 541         struct page *page = req->wb_page;
 542         struct nfs_lock_context *l_ctx = req->wb_lock_context;
 543         struct nfs_open_context *ctx;
 544
 545         if (page != NULL) {
 546                 put_page(page);
 547                 req->wb_page = NULL;
 548         }
 549         if (l_ctx != NULL) {
 550                 if (atomic_dec_and_test(&l_ctx->io_count)) {
 551                         wake_up_var(&l_ctx->io_count);
 552                         ctx = l_ctx->open_context;
 553                         if (test_bit(NFS_CONTEXT_UNLOCK, &ctx->flags))
 554                                 rpc_wake_up(&NFS_SERVER(d_inode(ctx->dentry))->uoc_rpcwaitq);
 555                 }
 556                 nfs_put_lock_context(l_ctx);
 557                 req->wb_lock_context = NULL;
 558         }
 559 }
 560
 561 /**
 562  * nfs_release_request - Release the count on an NFS read/write request
 563  * @req: request to release
 564  *
 565  * Note: Should never be called with the spinlock held!
 566  */
 567 void nfs_free_request(struct nfs_page *req)
 568 {
 569         WARN_ON_ONCE(req->wb_this_page != req);
 570
 571         /* extra debug: make sure no sync bits are still set */
 572         WARN_ON_ONCE(test_bit(PG_TEARDOWN, &req->wb_flags));
 573         WARN_ON_ONCE(test_bit(PG_UNLOCKPAGE, &req->wb_flags));
 574         WARN_ON_ONCE(test_bit(PG_UPTODATE, &req->wb_flags));
 575         WARN_ON_ONCE(test_bit(PG_WB_END, &req->wb_flags));
 576         WARN_ON_ONCE(test_bit(PG_REMOVE, &req->wb_flags));
 577
 578         /* Release struct file and open context */
 579         nfs_clear_request(req);
 580         nfs_page_free(req);
 581 }
 582
 583 void nfs_release_request(struct nfs_page *req)
 584 {
 585         kref_put(&req->wb_kref, nfs_page_group_destroy);
 586 }
 587 EXPORT_SYMBOL_GPL(nfs_release_request);
 588
 589 /**
 590  * nfs_wait_on_request - Wait for a request to complete.
 591  * @req: request to wait upon.
 592  *
 593  * Interruptible by fatal signals only.
 594  * The user is responsible for holding a count on the request.
 595  */
 596 int
 597 nfs_wait_on_request(struct nfs_page *req)
 598 {
 599         if (!test_bit(PG_BUSY, &req->wb_flags))
 600                 return 0;
 601         set_bit(PG_CONTENDED2, &req->wb_flags);
 602         smp_mb__after_atomic();
 603         return wait_on_bit_io(&req->wb_flags, PG_BUSY,
 604                               TASK_UNINTERRUPTIBLE);
 605 }
 606 EXPORT_SYMBOL_GPL(nfs_wait_on_request);
 607
 608 /*
 609  * nfs_generic_pg_test - determine if requests can be coalesced
 610  * @desc: pointer to descriptor
 611  * @prev: previous request in desc, or NULL
 612  * @req: this request
 613  *
 614  * Returns zero if @req cannot be coalesced into @desc, otherwise it returns
 615  * the size of the request.
 616  */
 617 size_t nfs_generic_pg_test(struct nfs_pageio_descriptor *desc,
 618                            struct nfs_page *prev, struct nfs_page *req)
 619 {
 620         struct nfs_pgio_mirror *mirror = nfs_pgio_current_mirror(desc);
 621
 622
 623         if (mirror->pg_count > mirror->pg_bsize) {
 624                 /* should never happen */
 625                 WARN_ON_ONCE(1);
 626                 return 0;
 627         }
 628
 629         /*
 630          * Limit the request size so that we can still allocate a page array
 631          * for it without upsetting the slab allocator.
 632          */
 633         if (((mirror->pg_count + req->wb_bytes) >> PAGE_SHIFT) *
 634                         sizeof(struct page *) > PAGE_SIZE)
 635                 return 0;
 636
 637         return min(mirror->pg_bsize - mirror->pg_count, (size_t)req->wb_bytes);
 638 }
 639 EXPORT_SYMBOL_GPL(nfs_generic_pg_test);
 640
 641 struct nfs_pgio_header *nfs_pgio_header_alloc(const struct nfs_rw_ops *ops)
 642 {
 643         struct nfs_pgio_header *hdr = ops->rw_alloc_header();
 644
 645         if (hdr) {
 646                 INIT_LIST_HEAD(&hdr->pages);
 647                 hdr->rw_ops = ops;
 648         }
 649         return hdr;
 650 }
 651 EXPORT_SYMBOL_GPL(nfs_pgio_header_alloc);
 652
 653 /**
 654  * nfs_pgio_data_destroy - make @hdr suitable for reuse
 655  *
 656  * Frees memory and releases refs from nfs_generic_pgio, so that it may
 657  * be called again.
 658  *
 659  * @hdr: A header that has had nfs_generic_pgio called
 660  */
 661 static void nfs_pgio_data_destroy(struct nfs_pgio_header *hdr)
 662 {
 663         if (hdr->args.context)
 664                 put_nfs_open_context(hdr->args.context);
 665         if (hdr->page_array.pagevec != hdr->page_array.page_array)
 666                 kfree(hdr->page_array.pagevec);
 667 }
 668
 669 /*
 670  * nfs_pgio_header_free - Free a read or write header
 671  * @hdr: The header to free
 672  */
 673 void nfs_pgio_header_free(struct nfs_pgio_header *hdr)
 674 {
 675         nfs_pgio_data_destroy(hdr);
 676         hdr->rw_ops->rw_free_header(hdr);
 677 }
 678 EXPORT_SYMBOL_GPL(nfs_pgio_header_free);
 679
 680 /**
 681  * nfs_pgio_rpcsetup - Set up arguments for a pageio call
 682  * @hdr: The pageio hdr
 683  * @count: Number of bytes to read
 684  * @how: How to commit data (writes only)
 685  * @cinfo: Commit information for the call (writes only)
 686  */
 687 static void nfs_pgio_rpcsetup(struct nfs_pgio_header *hdr,
 688                               unsigned int count,
 689                               int how, struct nfs_commit_info *cinfo)
 690 {
 691         struct nfs_page *req = hdr->req;
 692
 693         /* Set up the RPC argument and reply structs
 694          * NB: take care not to mess about with hdr->commit et al. */
 695
 696         hdr->args.fh     = NFS_FH(hdr->inode);
 697         hdr->args.offset = req_offset(req);
 698         /* pnfs_set_layoutcommit needs this */
 699         hdr->mds_offset = hdr->args.offset;
 700         hdr->args.pgbase = req->wb_pgbase;
 701         hdr->args.pages  = hdr->page_array.pagevec;
 702         hdr->args.count  = count;
 703         hdr->args.context = get_nfs_open_context(nfs_req_openctx(req));
 704         hdr->args.lock_context = req->wb_lock_context;
 705         hdr->args.stable  = NFS_UNSTABLE;
 706         switch (how & (FLUSH_STABLE | FLUSH_COND_STABLE)) {
 707         case 0:
 708                 break;
 709         case FLUSH_COND_STABLE:
 710                 if (nfs_reqs_to_commit(cinfo))
 711                         break;
 712                 /* fall through */
 713         default:
 714                 hdr->args.stable = NFS_FILE_SYNC;
 715         }
 716
 717         hdr->res.fattr   = &hdr->fattr;
 718         hdr->res.count   = 0;
 719         hdr->res.eof     = 0;
 720         hdr->res.verf    = &hdr->verf;
 721         nfs_fattr_init(&hdr->fattr);
 722 }
 723
 724 /**
 725  * nfs_pgio_prepare - Prepare pageio hdr to go over the wire
 726  * @task: The current task
 727  * @calldata: pageio header to prepare
 728  */
 729 static void nfs_pgio_prepare(struct rpc_task *task, void *calldata)
 730 {
 731         struct nfs_pgio_header *hdr = calldata;
 732         int err;
 733         err = NFS_PROTO(hdr->inode)->pgio_rpc_prepare(task, hdr);
 734         if (err)
 735                 rpc_exit(task, err);
 736 }
 737
 738 int nfs_initiate_pgio(struct rpc_clnt *clnt, struct nfs_pgio_header *hdr,
 739                       const struct cred *cred, const struct nfs_rpc_ops *rpc_ops,
 740                       const struct rpc_call_ops *call_ops, int how, int flags)
 741 {
 742         struct rpc_task *task;
 743         struct rpc_message msg = {
 744                 .rpc_argp = &hdr->args,
 745                 .rpc_resp = &hdr->res,
 746                 .rpc_cred = cred,
 747         };
 748         struct rpc_task_setup task_setup_data = {
 749                 .rpc_client = clnt,
 750                 .task = &hdr->task,
 751                 .rpc_message = &msg,
 752                 .callback_ops = call_ops,
 753                 .callback_data = hdr,
 754                 .workqueue = nfsiod_workqueue,
 755                 .flags = RPC_TASK_ASYNC | flags,
 756         };
 757
 758         hdr->rw_ops->rw_initiate(hdr, &msg, rpc_ops, &task_setup_data, how);
 759
 760         dprintk("NFS: initiated pgio call "
 761                 "(req %s/%llu, %u bytes @ offset %llu)\n",
 762                 hdr->inode->i_sb->s_id,
 763                 (unsigned long long)NFS_FILEID(hdr->inode),
 764                 hdr->args.count,
 765                 (unsigned long long)hdr->args.offset);
 766
 767         task = rpc_run_task(&task_setup_data);
 768         if (IS_ERR(task))
 769                 return PTR_ERR(task);
 770         rpc_put_task(task);
 771         return 0;
 772 }
 773 EXPORT_SYMBOL_GPL(nfs_initiate_pgio);
 774
 775 /**
 776  * nfs_pgio_error - Clean up from a pageio error
 777  * @hdr: pageio header
 778  */
 779 static void nfs_pgio_error(struct nfs_pgio_header *hdr)
 780 {
 781         set_bit(NFS_IOHDR_REDO, &hdr->flags);
 782         hdr->completion_ops->completion(hdr);
 783 }
 784
 785 /**
 786  * nfs_pgio_release - Release pageio data
 787  * @calldata: The pageio header to release
 788  */
 789 static void nfs_pgio_release(void *calldata)
 790 {
 791         struct nfs_pgio_header *hdr = calldata;
 792         hdr->completion_ops->completion(hdr);
 793 }
 794
 795 static void nfs_pageio_mirror_init(struct nfs_pgio_mirror *mirror,
 796                                    unsigned int bsize)
 797 {
 798         INIT_LIST_HEAD(&mirror->pg_list);
 799         mirror->pg_bytes_written = 0;
 800         mirror->pg_count = 0;
 801         mirror->pg_bsize = bsize;
 802         mirror->pg_base = 0;
 803         mirror->pg_recoalesce = 0;
 804 }
 805
 806 /**
 807  * nfs_pageio_init - initialise a page io descriptor
 808  * @desc: pointer to descriptor
 809  * @inode: pointer to inode
 810  * @pg_ops: pointer to pageio operations
 811  * @compl_ops: pointer to pageio completion operations
 812  * @rw_ops: pointer to nfs read/write operations
 813  * @bsize: io block size
 814  * @io_flags: extra parameters for the io function
 815  */
 816 void nfs_pageio_init(struct nfs_pageio_descriptor *desc,
 817                      struct inode *inode,
 818                      const struct nfs_pageio_ops *pg_ops,
 819                      const struct nfs_pgio_completion_ops *compl_ops,
 820                      const struct nfs_rw_ops *rw_ops,
 821                      size_t bsize,
 822                      int io_flags)
 823 {
 824         desc->pg_moreio = 0;
 825         desc->pg_inode = inode;
 826         desc->pg_ops = pg_ops;
 827         desc->pg_completion_ops = compl_ops;
 828         desc->pg_rw_ops = rw_ops;
 829         desc->pg_ioflags = io_flags;
 830         desc->pg_error = 0;
 831         desc->pg_lseg = NULL;
 832         desc->pg_io_completion = NULL;
 833         desc->pg_dreq = NULL;
 834         desc->pg_bsize = bsize;
 835
 836         desc->pg_mirror_count = 1;
 837         desc->pg_mirror_idx = 0;
 838
 839         desc->pg_mirrors_dynamic = NULL;
 840         desc->pg_mirrors = desc->pg_mirrors_static;
 841         nfs_pageio_mirror_init(&desc->pg_mirrors[0], bsize);
 842         desc->pg_maxretrans = 0;
 843 }
 844
 845 /**
 846  * nfs_pgio_result - Basic pageio error handling
 847  * @task: The task that ran
 848  * @calldata: Pageio header to check
 849  */
 850 static void nfs_pgio_result(struct rpc_task *task, void *calldata)
 851 {
 852         struct nfs_pgio_header *hdr = calldata;
 853         struct inode *inode = hdr->inode;
 854
 855         dprintk("NFS: %s: %5u, (status %d)\n", __func__,
 856                 task->tk_pid, task->tk_status);
 857
 858         if (hdr->rw_ops->rw_done(task, hdr, inode) != 0)
 859                 return;
 860         if (task->tk_status < 0)
 861                 nfs_set_pgio_error(hdr, task->tk_status, hdr->args.offset);
 862         else
 863                 hdr->rw_ops->rw_result(task, hdr);
 864 }
 865
 866 /*
 867  * Create an RPC task for the given read or write request and kick it.
 868  * The page must have been locked by the caller.
 869  *
 870  * It may happen that the page we're passed is not marked dirty.
 871  * This is the case if nfs_updatepage detects a conflicting request
 872  * that has been written but not committed.
 873  */
 874 int nfs_generic_pgio(struct nfs_pageio_descriptor *desc,
 875                      struct nfs_pgio_header *hdr)
 876 {
 877         struct nfs_pgio_mirror *mirror = nfs_pgio_current_mirror(desc);
 878
 879         struct nfs_page         *req;
 880         struct page             **pages,
 881                                 *last_page;
 882         struct list_head *head = &mirror->pg_list;
 883         struct nfs_commit_info cinfo;
 884         struct nfs_page_array *pg_array = &hdr->page_array;
 885         unsigned int pagecount, pageused;
 886         gfp_t gfp_flags = GFP_KERNEL;
 887
 888         pagecount = nfs_page_array_len(mirror->pg_base, mirror->pg_count);
 889         pg_array->npages = pagecount;
 890
 891         if (pagecount <= ARRAY_SIZE(pg_array->page_array))
 892                 pg_array->pagevec = pg_array->page_array;
 893         else {
 894                 pg_array->pagevec = kcalloc(pagecount, sizeof(struct page *), gfp_flags);
 895                 if (!pg_array->pagevec) {
 896                         pg_array->npages = 0;
 897                         nfs_pgio_error(hdr);
 898                         desc->pg_error = -ENOMEM;
 899                         return desc->pg_error;
 900                 }
 901         }
 902
 903         nfs_init_cinfo(&cinfo, desc->pg_inode, desc->pg_dreq);
 904         pages = hdr->page_array.pagevec;
 905         last_page = NULL;
 906         pageused = 0;
 907         while (!list_empty(head)) {
 908                 req = nfs_list_entry(head->next);
 909                 nfs_list_move_request(req, &hdr->pages);
 910
 911                 if (!last_page || last_page != req->wb_page) {
 912                         pageused++;
 913                         if (pageused > pagecount)
 914                                 break;
 915                         *pages++ = last_page = req->wb_page;
 916                 }
 917         }
 918         if (WARN_ON_ONCE(pageused != pagecount)) {
 919                 nfs_pgio_error(hdr);
 920                 desc->pg_error = -EINVAL;
 921                 return desc->pg_error;
 922         }
 923
 924         if ((desc->pg_ioflags & FLUSH_COND_STABLE) &&
 925             (desc->pg_moreio || nfs_reqs_to_commit(&cinfo)))
 926                 desc->pg_ioflags &= ~FLUSH_COND_STABLE;
 927
 928         /* Set up the argument struct */
 929         nfs_pgio_rpcsetup(hdr, mirror->pg_count, desc->pg_ioflags, &cinfo);
 930         desc->pg_rpc_callops = &nfs_pgio_common_ops;
 931         return 0;
 932 }
 933 EXPORT_SYMBOL_GPL(nfs_generic_pgio);
 934
 935 static int nfs_generic_pg_pgios(struct nfs_pageio_descriptor *desc)
 936 {
 937         struct nfs_pgio_header *hdr;
 938         int ret;
 939
 940         hdr = nfs_pgio_header_alloc(desc->pg_rw_ops);
 941         if (!hdr) {
 942                 desc->pg_error = -ENOMEM;
 943                 return desc->pg_error;
 944         }
 945         nfs_pgheader_init(desc, hdr, nfs_pgio_header_free);
 946         ret = nfs_generic_pgio(desc, hdr);
 947         if (ret == 0)
 948                 ret = nfs_initiate_pgio(NFS_CLIENT(hdr->inode),
 949                                         hdr,
 950                                         hdr->cred,
 951                                         NFS_PROTO(hdr->inode),
 952                                         desc->pg_rpc_callops,
 953                                         desc->pg_ioflags,
 954                                         RPC_TASK_CRED_NOREF);
 955         return ret;
 956 }
 957
 958 static struct nfs_pgio_mirror *
 959 nfs_pageio_alloc_mirrors(struct nfs_pageio_descriptor *desc,
 960                 unsigned int mirror_count)
 961 {
 962         struct nfs_pgio_mirror *ret;
 963         unsigned int i;
 964
 965         kfree(desc->pg_mirrors_dynamic);
 966         desc->pg_mirrors_dynamic = NULL;
 967         if (mirror_count == 1)
 968                 return desc->pg_mirrors_static;
 969         ret = kmalloc_array(mirror_count, sizeof(*ret), GFP_KERNEL);
 970         if (ret != NULL) {
 971                 for (i = 0; i < mirror_count; i++)
 972                         nfs_pageio_mirror_init(&ret[i], desc->pg_bsize);
 973                 desc->pg_mirrors_dynamic = ret;
 974         }
 975         return ret;
 976 }
 977
 978 /*
 979  * nfs_pageio_setup_mirroring - determine if mirroring is to be used
 980  *                              by calling the pg_get_mirror_count op
 981  */
 982 static void nfs_pageio_setup_mirroring(struct nfs_pageio_descriptor *pgio,
 983                                        struct nfs_page *req)
 984 {
 985         unsigned int mirror_count = 1;
 986
 987         if (pgio->pg_ops->pg_get_mirror_count)
 988                 mirror_count = pgio->pg_ops->pg_get_mirror_count(pgio, req);
 989         if (mirror_count == pgio->pg_mirror_count || pgio->pg_error < 0)
 990                 return;
 991
 992         if (!mirror_count || mirror_count > NFS_PAGEIO_DESCRIPTOR_MIRROR_MAX) {
 993                 pgio->pg_error = -EINVAL;
 994                 return;
 995         }
 996
 997         pgio->pg_mirrors = nfs_pageio_alloc_mirrors(pgio, mirror_count);
 998         if (pgio->pg_mirrors == NULL) {
 999                 pgio->pg_error = -ENOMEM;
1000                 pgio->pg_mirrors = pgio->pg_mirrors_static;
1001                 mirror_count = 1;
1002         }
1003         pgio->pg_mirror_count = mirror_count;
1004 }
1005
1006 static void nfs_pageio_cleanup_mirroring(struct nfs_pageio_descriptor *pgio)
1007 {
1008         pgio->pg_mirror_count = 1;
1009         pgio->pg_mirror_idx = 0;
1010         pgio->pg_mirrors = pgio->pg_mirrors_static;
1011         kfree(pgio->pg_mirrors_dynamic);
1012         pgio->pg_mirrors_dynamic = NULL;
1013 }
1014
1015 static bool nfs_match_lock_context(const struct nfs_lock_context *l1,
1016                 const struct nfs_lock_context *l2)
1017 {
1018         return l1->lockowner == l2->lockowner;
1019 }
1020
1021 /**
1022  * nfs_coalesce_size - test two requests for compatibility
1023  * @prev: pointer to nfs_page
1024  * @req: pointer to nfs_page
1025  * @pgio: pointer to nfs_pagio_descriptor
1026  *
1027  * The nfs_page structures 'prev' and 'req' are compared to ensure that the
1028  * page data area they describe is contiguous, and that their RPC
1029  * credentials, NFSv4 open state, and lockowners are the same.
1030  *
1031  * Returns size of the request that can be coalesced
1032  */
1033 static unsigned int nfs_coalesce_size(struct nfs_page *prev,
1034                                       struct nfs_page *req,
1035                                       struct nfs_pageio_descriptor *pgio)
1036 {
1037         struct file_lock_context *flctx;
1038
1039         if (prev) {
1040                 if (!nfs_match_open_context(nfs_req_openctx(req), nfs_req_openctx(prev)))
1041                         return 0;
1042                 flctx = d_inode(nfs_req_openctx(req)->dentry)->i_flctx;
1043                 if (flctx != NULL &&
1044                     !(list_empty_careful(&flctx->flc_posix) &&
1045                       list_empty_careful(&flctx->flc_flock)) &&
1046                     !nfs_match_lock_context(req->wb_lock_context,
1047                                             prev->wb_lock_context))
1048                         return 0;
1049                 if (req_offset(req) != req_offset(prev) + prev->wb_bytes)
1050                         return 0;
1051                 if (req->wb_page == prev->wb_page) {
1052                         if (req->wb_pgbase != prev->wb_pgbase + prev->wb_bytes)
1053                                 return 0;
1054                 } else {
1055                         if (req->wb_pgbase != 0 ||
1056                             prev->wb_pgbase + prev->wb_bytes != PAGE_SIZE)
1057                                 return 0;
1058                 }
1059         }
1060         return pgio->pg_ops->pg_test(pgio, prev, req);
1061 }
1062
1063 /**
1064  * nfs_pageio_do_add_request - Attempt to coalesce a request into a page list.
1065  * @desc: destination io descriptor
1066  * @req: request
1067  *
1068  * If the request 'req' was successfully coalesced into the existing list
1069  * of pages 'desc', it returns the size of req.
1070  */
1071 static unsigned int
1072 nfs_pageio_do_add_request(struct nfs_pageio_descriptor *desc,
1073                 struct nfs_page *req)
1074 {
1075         struct nfs_pgio_mirror *mirror = nfs_pgio_current_mirror(desc);
1076         struct nfs_page *prev = NULL;
1077         unsigned int size;
1078
1079         if (mirror->pg_count != 0) {
1080                 prev = nfs_list_entry(mirror->pg_list.prev);
1081         } else {
1082                 if (desc->pg_ops->pg_init)
1083                         desc->pg_ops->pg_init(desc, req);
1084                 if (desc->pg_error < 0)
1085                         return 0;
1086                 mirror->pg_base = req->wb_pgbase;
1087         }
1088
1089         if (desc->pg_maxretrans && req->wb_nio > desc->pg_maxretrans) {
1090                 if (NFS_SERVER(desc->pg_inode)->flags & NFS_MOUNT_SOFTERR)
1091                         desc->pg_error = -ETIMEDOUT;
1092                 else
1093                         desc->pg_error = -EIO;
1094                 return 0;
1095         }
1096
1097         size = nfs_coalesce_size(prev, req, desc);
1098         if (size < req->wb_bytes)
1099                 return size;
1100         nfs_list_move_request(req, &mirror->pg_list);
1101         mirror->pg_count += req->wb_bytes;
1102         return req->wb_bytes;
1103 }
1104
1105 /*
1106  * Helper for nfs_pageio_add_request and nfs_pageio_complete
1107  */
1108 static void nfs_pageio_doio(struct nfs_pageio_descriptor *desc)
1109 {
1110         struct nfs_pgio_mirror *mirror = nfs_pgio_current_mirror(desc);
1111
1112
1113         if (!list_empty(&mirror->pg_list)) {
1114                 int error = desc->pg_ops->pg_doio(desc);
1115                 if (error < 0)
1116                         desc->pg_error = error;
1117                 else
1118                         mirror->pg_bytes_written += mirror->pg_count;
1119         }
1120         if (list_empty(&mirror->pg_list)) {
1121                 mirror->pg_count = 0;
1122                 mirror->pg_base = 0;
1123         }
1124 }
1125
1126 static void
1127 nfs_pageio_cleanup_request(struct nfs_pageio_descriptor *desc,
1128                 struct nfs_page *req)
1129 {
1130         LIST_HEAD(head);
1131
1132         nfs_list_move_request(req, &head);
1133         desc->pg_completion_ops->error_cleanup(&head, desc->pg_error);
1134 }
1135
1136 /**
1137  * nfs_pageio_add_request - Attempt to coalesce a request into a page list.
1138  * @desc: destination io descriptor
1139  * @req: request
1140  *
1141  * This may split a request into subrequests which are all part of the
1142  * same page group. If so, it will submit @req as the last one, to ensure
1143  * the pointer to @req is still valid in case of failure.
1144  *
1145  * Returns true if the request 'req' was successfully coalesced into the
1146  * existing list of pages 'desc'.
1147  */
1148 static int __nfs_pageio_add_request(struct nfs_pageio_descriptor *desc,
1149                            struct nfs_page *req)
1150 {
1151         struct nfs_pgio_mirror *mirror = nfs_pgio_current_mirror(desc);
1152         struct nfs_page *subreq;
1153         unsigned int size, subreq_size;
1154
1155         nfs_page_group_lock(req);
1156
1157         subreq = req;
1158         subreq_size = subreq->wb_bytes;
1159         for(;;) {
1160                 size = nfs_pageio_do_add_request(desc, subreq);
1161                 if (size == subreq_size) {
1162                         /* We successfully submitted a request */
1163                         if (subreq == req)
1164                                 break;
1165                         req->wb_pgbase += size;
1166                         req->wb_bytes -= size;
1167                         req->wb_offset += size;
1168                         subreq_size = req->wb_bytes;
1169                         subreq = req;
1170                         continue;
1171                 }
1172                 if (WARN_ON_ONCE(subreq != req)) {
1173                         nfs_page_group_unlock(req);
1174                         nfs_pageio_cleanup_request(desc, subreq);
1175                         subreq = req;
1176                         subreq_size = req->wb_bytes;
1177                         nfs_page_group_lock(req);
1178                 }
1179                 if (!size) {
1180                         /* Can't coalesce any more, so do I/O */
1181                         nfs_page_group_unlock(req);
1182                         desc->pg_moreio = 1;
1183                         nfs_pageio_doio(desc);
1184                         if (desc->pg_error < 0 || mirror->pg_recoalesce)
1185                                 return 0;
1186                         /* retry add_request for this subreq */
1187                         nfs_page_group_lock(req);
1188                         continue;
1189                 }
1190                 subreq = nfs_create_subreq(req, req->wb_pgbase,
1191                                 req->wb_offset, size);
1192                 if (IS_ERR(subreq))
1193                         goto err_ptr;
1194                 subreq_size = size;
1195         }
1196
1197         nfs_page_group_unlock(req);
1198         return 1;
1199 err_ptr:
1200         desc->pg_error = PTR_ERR(subreq);
1201         nfs_page_group_unlock(req);
1202         return 0;
1203 }
1204
1205 static int nfs_do_recoalesce(struct nfs_pageio_descriptor *desc)
1206 {
1207         struct nfs_pgio_mirror *mirror = nfs_pgio_current_mirror(desc);
1208         LIST_HEAD(head);
1209
1210         do {
1211                 list_splice_init(&mirror->pg_list, &head);
1212                 mirror->pg_bytes_written -= mirror->pg_count;
1213                 mirror->pg_count = 0;
1214                 mirror->pg_base = 0;
1215                 mirror->pg_recoalesce = 0;
1216
1217                 while (!list_empty(&head)) {
1218                         struct nfs_page *req;
1219
1220                         req = list_first_entry(&head, struct nfs_page, wb_list);
1221                         if (__nfs_pageio_add_request(desc, req))
1222                                 continue;
1223                         if (desc->pg_error < 0) {
1224                                 list_splice_tail(&head, &mirror->pg_list);
1225                                 mirror->pg_recoalesce = 1;
1226                                 return 0;
1227                         }
1228                         break;
1229                 }
1230         } while (mirror->pg_recoalesce);
1231         return 1;
1232 }
1233
1234 static int nfs_pageio_add_request_mirror(struct nfs_pageio_descriptor *desc,
1235                 struct nfs_page *req)
1236 {
1237         int ret;
1238
1239         do {
1240                 ret = __nfs_pageio_add_request(desc, req);
1241                 if (ret)
1242                         break;
1243                 if (desc->pg_error < 0)
1244                         break;
1245                 ret = nfs_do_recoalesce(desc);
1246         } while (ret);
1247
1248         return ret;
1249 }
1250
1251 static void nfs_pageio_error_cleanup(struct nfs_pageio_descriptor *desc)
1252 {
1253         u32 midx;
1254         struct nfs_pgio_mirror *mirror;
1255
1256         if (!desc->pg_error)
1257                 return;
1258
1259         for (midx = 0; midx < desc->pg_mirror_count; midx++) {
1260                 mirror = &desc->pg_mirrors[midx];
1261                 desc->pg_completion_ops->error_cleanup(&mirror->pg_list,
1262                                 desc->pg_error);
1263         }
1264 }
1265
1266 int nfs_pageio_add_request(struct nfs_pageio_descriptor *desc,
1267                            struct nfs_page *req)
1268 {
1269         u32 midx;
1270         unsigned int pgbase, offset, bytes;
1271         struct nfs_page *dupreq;
1272
1273         pgbase = req->wb_pgbase;
1274         offset = req->wb_offset;
1275         bytes = req->wb_bytes;
1276
1277         nfs_pageio_setup_mirroring(desc, req);
1278         if (desc->pg_error < 0)
1279                 goto out_failed;
1280
1281         /* Create the mirror instances first, and fire them off */
1282         for (midx = 1; midx < desc->pg_mirror_count; midx++) {
1283                 nfs_page_group_lock(req);
1284
1285                 dupreq = nfs_create_subreq(req,
1286                                 pgbase, offset, bytes);
1287
1288                 nfs_page_group_unlock(req);
1289                 if (IS_ERR(dupreq)) {
1290                         desc->pg_error = PTR_ERR(dupreq);
1291                         goto out_failed;
1292                 }
1293
1294                 desc->pg_mirror_idx = midx;
1295                 if (!nfs_pageio_add_request_mirror(desc, dupreq))
1296                         goto out_cleanup_subreq;
1297         }
1298
1299         desc->pg_mirror_idx = 0;
1300         if (!nfs_pageio_add_request_mirror(desc, req))
1301                 goto out_failed;
1302
1303         return 1;
1304
1305 out_cleanup_subreq:
1306         nfs_pageio_cleanup_request(desc, dupreq);
1307 out_failed:
1308         nfs_pageio_error_cleanup(desc);
1309         return 0;
1310 }
1311
1312 /*
1313  * nfs_pageio_complete_mirror - Complete I/O on the current mirror of an
1314  *                              nfs_pageio_descriptor
1315  * @desc: pointer to io descriptor
1316  * @mirror_idx: pointer to mirror index
1317  */
1318 static void nfs_pageio_complete_mirror(struct nfs_pageio_descriptor *desc,
1319                                        u32 mirror_idx)
1320 {
1321         struct nfs_pgio_mirror *mirror = &desc->pg_mirrors[mirror_idx];
1322         u32 restore_idx = desc->pg_mirror_idx;
1323
1324         desc->pg_mirror_idx = mirror_idx;
1325         for (;;) {
1326                 nfs_pageio_doio(desc);
1327                 if (desc->pg_error < 0 || !mirror->pg_recoalesce)
1328                         break;
1329                 if (!nfs_do_recoalesce(desc))
1330                         break;
1331         }
1332         desc->pg_mirror_idx = restore_idx;
1333 }
1334
1335 /*
1336  * nfs_pageio_resend - Transfer requests to new descriptor and resend
1337  * @hdr - the pgio header to move request from
1338  * @desc - the pageio descriptor to add requests to
1339  *
1340  * Try to move each request (nfs_page) from @hdr to @desc then attempt
1341  * to send them.
1342  *
1343  * Returns 0 on success and < 0 on error.
1344  */
1345 int nfs_pageio_resend(struct nfs_pageio_descriptor *desc,
1346                       struct nfs_pgio_header *hdr)
1347 {
1348         LIST_HEAD(pages);
1349
1350         desc->pg_io_completion = hdr->io_completion;
1351         desc->pg_dreq = hdr->dreq;
1352         list_splice_init(&hdr->pages, &pages);
1353         while (!list_empty(&pages)) {
1354                 struct nfs_page *req = nfs_list_entry(pages.next);
1355
1356                 if (!nfs_pageio_add_request(desc, req))
1357                         break;
1358         }
1359         nfs_pageio_complete(desc);
1360         if (!list_empty(&pages)) {
1361                 int err = desc->pg_error < 0 ? desc->pg_error : -EIO;
1362                 hdr->completion_ops->error_cleanup(&pages, err);
1363                 nfs_set_pgio_error(hdr, err, hdr->io_start);
1364                 return err;
1365         }
1366         return 0;
1367 }
1368 EXPORT_SYMBOL_GPL(nfs_pageio_resend);
1369
1370 /**
1371  * nfs_pageio_complete - Complete I/O then cleanup an nfs_pageio_descriptor
1372  * @desc: pointer to io descriptor
1373  */
1374 void nfs_pageio_complete(struct nfs_pageio_descriptor *desc)
1375 {
1376         u32 midx;
1377
1378         for (midx = 0; midx < desc->pg_mirror_count; midx++)
1379                 nfs_pageio_complete_mirror(desc, midx);
1380
1381         if (desc->pg_error < 0)
1382                 nfs_pageio_error_cleanup(desc);
1383         if (desc->pg_ops->pg_cleanup)
1384                 desc->pg_ops->pg_cleanup(desc);
1385         nfs_pageio_cleanup_mirroring(desc);
1386 }
1387
1388 /**
1389  * nfs_pageio_cond_complete - Conditional I/O completion
1390  * @desc: pointer to io descriptor
1391  * @index: page index
1392  *
1393  * It is important to ensure that processes don't try to take locks
1394  * on non-contiguous ranges of pages as that might deadlock. This
1395  * function should be called before attempting to wait on a locked
1396  * nfs_page. It will complete the I/O if the page index 'index'
1397  * is not contiguous with the existing list of pages in 'desc'.
1398  */
1399 void nfs_pageio_cond_complete(struct nfs_pageio_descriptor *desc, pgoff_t index)
1400 {
1401         struct nfs_pgio_mirror *mirror;
1402         struct nfs_page *prev;
1403         u32 midx;
1404
1405         for (midx = 0; midx < desc->pg_mirror_count; midx++) {
1406                 mirror = &desc->pg_mirrors[midx];
1407                 if (!list_empty(&mirror->pg_list)) {
1408                         prev = nfs_list_entry(mirror->pg_list.prev);
1409                         if (index != prev->wb_index + 1) {
1410                                 nfs_pageio_complete(desc);
1411                                 break;
1412                         }
1413                 }
1414         }
1415 }
1416
1417 /*
1418  * nfs_pageio_stop_mirroring - stop using mirroring (set mirror count to 1)
1419  */
1420 void nfs_pageio_stop_mirroring(struct nfs_pageio_descriptor *pgio)
1421 {
1422         nfs_pageio_complete(pgio);
1423 }
1424
1425 int __init nfs_init_nfspagecache(void)
1426 {
1427         nfs_page_cachep = kmem_cache_create("nfs_page",
1428                                             sizeof(struct nfs_page),
1429                                             0, SLAB_HWCACHE_ALIGN,
1430                                             NULL);
1431         if (nfs_page_cachep == NULL)
1432                 return -ENOMEM;
1433
1434         return 0;
1435 }
1436
1437 void nfs_destroy_nfspagecache(void)
1438 {
1439         kmem_cache_destroy(nfs_page_cachep);
1440 }
1441
1442 static const struct rpc_call_ops nfs_pgio_common_ops = {
1443         .rpc_call_prepare = nfs_pgio_prepare,
1444         .rpc_call_done = nfs_pgio_result,
1445         .rpc_release = nfs_pgio_release,
1446 };
1447
1448 const struct nfs_pageio_ops nfs_pgio_rw_ops = {
1449         .pg_test = nfs_generic_pg_test,
1450         .pg_doio = nfs_generic_pg_pgios,
1451 };